Contextune Delegation Modes¶

Contextune now supports three delegation modes that control how detected commands are handled, including a sub-agent verification mode that preserves your main agent's context.

Overview¶

Mode	Behavior	Context Usage	Safety	Speed
verify	Spawns sub-agent to ask user	Minimal	✅ Safest	⚡⚡ Fast
suggest	Adds context, main agent decides	Moderate	✅ Safe	⚡⚡ Fast
auto	Auto-executes detected command	None	⚠️ Risky	⚡⚡⚡ Fastest

Verify Mode (Recommended)¶

How it works:

User types: "analyze my code"
    ↓
Hook detects: /sc:analyze
    ↓
Main agent receives delegation directive
    ↓
Main agent spawns sub-agent
    ↓
Sub-agent asks user: "Option 1: /sc:analyze or Option 2: original prompt?"
    ↓
Sub-agent executes user's choice
    ↓
Sub-agent reports back to main agent
    ↓
Main agent context preserved!

Benefits: - ✅ User verification - You confirm before execution - ✅ Context preservation - Main agent doesn't waste tokens - ✅ Safe - No unexpected command execution - ✅ Flexible - You can choose original prompt or detected command

Example Flow:

You: "can you help me analyze this code for bugs"

Claude (Main): 🎯 Contextune detected a command. Spawning verification agent...

[Sub-agent spawned]

Claude (Sub-agent):
🎯 **Contextune Detection**

I detected you might want to run a slash command:

**Option 1**: Run `/sc:analyze` (85% confidence)
**Option 2**: Continue with original: "can you help me analyze this code for bugs"

Which do you prefer? (1, 2, or tell me what you want)

You: 1

Claude (Sub-agent):
✅ Executing /sc:analyze...

[Performs analysis]

**Contextune Verification Complete**
User choice: Option 1 (/sc:analyze)
Action taken: Ran code analysis
Result: Found 3 potential issues

[Returns to main agent]

Claude (Main): [continues with minimal context consumed]

Suggest Mode¶

How it works:

User types: "analyze my code"
    ↓
Hook detects: /sc:analyze
    ↓
Main agent receives: original prompt + suggestion
    ↓
Main agent asks user which they prefer
    ↓
Main agent executes

Benefits: - ✅ User verification - ⚡ Slightly faster (no sub-agent spawn) - ⚠️ Uses main agent context for verification

When to use: - You want verification but don't care about context usage - Quick one-off tasks

Auto Mode (Original Behavior)¶

How it works:

User types: "analyze my code"
    ↓
Hook detects: /sc:analyze
    ↓
Main agent receives: "/sc:analyze" (original prompt replaced!)
    ↓
Executes immediately

Benefits: - ⚡⚡⚡ Fastest - No interruption

Risks: - ❌ No verification - ❌ Original prompt context lost - ❌ Wrong detections execute immediately

When to use: - You trust the detection 100% - High-confidence scenarios only - You understand the risks

Configuration¶

Set Delegation Mode¶

Edit ~/.claude/plugins/contextune/data/user_patterns.json:

{
  "enabled": true,
  "confidence_threshold": 0.7,
  "delegation_mode": "verify",  // ← Change this
  "tiers": {
    "keyword": true,
    "model2vec": true,
    "semantic_router": true
  }
}

Options: - "verify" - Sub-agent verification (recommended) - "suggest" - Main agent asks - "auto" - Auto-execute

View Current Configuration¶

/contextune:config

Advanced: Hybrid Mode (Coming Soon)¶

Future version will support hybrid mode:

{
  "delegation_mode": "hybrid",
  "auto_execute_threshold": 0.95,
  "auto_execute_whitelist": [
    "/sc:analyze",
    "/sc:explain",
    "/contextune:stats"
  ]
}

Behavior: - Confidence ≥ 95% + whitelisted → auto-execute - Confidence 70-95% → verify mode - Confidence < 70% → pass through

Technical Details¶

Sub-Agent Architecture¶

Main Agent receives:

{
  "continue": true,
  "hookSpecificOutput": {
    "additionalContext": "[Delegation directive with detection details]"
  },
  "feedback": "🎯 Contextune: Spawning verification agent"
}

Main Agent spawns sub-agent with:

Task(
    subagent_type="general-purpose",
    description="Verify Contextune detection",
    prompt="[Verification instructions with detected command]"
)

Sub-agent: 1. Presents options to user 2. Waits for user choice 3. Executes chosen action 4. Reports results 5. Exits (context discarded)

Main agent: - Receives brief summary - Continues with preserved context

Context Savings¶

Without sub-agent delegation:

Main agent tokens: Original prompt (20) + Detection message (50) + User response (10) + Execution (200) = 280 tokens

With sub-agent delegation:

Main agent tokens: Delegation directive (100) + Sub-agent summary (50) = 150 tokens
Sub-agent tokens: Verification (50) + User response (10) + Execution (200) = 260 tokens
Total: 410 tokens (but main agent only uses 150)

Savings: Main agent uses 46% less context!

FAQ¶

Q: Can I change modes mid-conversation? A: Yes! Edit the config file and the next prompt will use the new mode.

Q: What if I don't want Contextune at all? A: Set "enabled": false in config, or uninstall: /plugin uninstall contextune

Q: Can the sub-agent access my files? A: Yes, sub-agents have the same permissions as the main agent.

Q: What if detection is wrong? A: In verify/suggest modes, you choose. In auto mode, you're stuck with it (use verify mode!)

Q: Does this work with all commands? A: Yes, Contextune detects any slash command it's trained on.

Recommendations¶

For safety: Use "verify" mode For speed: Use "auto" mode (but understand the risks) For balance: Use "suggest" mode

Default: "verify" mode is the safest and recommended option.